SAN JOSE, Calif. — Knowles Corp. is sampling an audio SoC to expand beyond its position in MEMS microphones. The company hopes to ride sockets in a wide variety of voice-enabled systems that it expects OEMs to ship over the next year.

Smart speakers such as the Amazon Echo are “just one form factor of many to come,” said Mike Polacek, a semiconductor veteran who joined Knowles earlier this year to head a new intelligent audio group.

“There will be a great deal of experimentation and innovation to determine what devices need voice compared to buttons or switches and at what level … we are at the very early stages.”

The company sees headsets as one significant emerging market along with experiments in everything from home appliances and thermostats to light switches and bulbs. Some products will only use voice to handle a few basic commands.

Leveraging its estimated 50 percent market share in MEMS microphones in handsets and other consumer devices, Knowles rolled out a mic with an embedded DSP earlier this year. The IA610 uses just 128 Kbytes of memory to handle basic tasks such as responding to a few wake words, but it “can’t do full speech recognition or respond to 100 wake words,” said Polacek.

The IA610 was the first fruit of Knowles’ acquisition in 2015 of Audience, a developer of audio algorithms on DSPs. Audience’s technology also fueled the new IA8508, an SoC supporting up to eight microphones, which Knowles hopes will lead to even bigger opportunities.

Three Tensilica DSPs lay at the heart of the 122-MHz chip, each optimized for different tasks. The HemiDelta is an ultra-low power, two-way SIMD floating-point DSP to keep the system on and listening for wake words; the DeltaMax is a four-way SIMD core optimized for performance; and the SSP targets low-latency jobs such as active noise cancellation in as little as 10 microseconds.

Next page: China cranks out audio algorithms

text

The IA8508 uses an ARM M4 and three optimized Tensilica DSPs. Click to enlarge. (Image: Knowles)

The DSPs use Tensilica’s Xtensa HiFi 3 instruction set and proprietary extensions from Knowles. As successful OEM experiments define new kinds of voice-enabled products, Knowles will target them with new custom instructions and other optimizations.

While DSPs deliver best audio performance and code efficiency, they can be hard to program. Thus, Knowles is working to get third-party algorithm developers to port code to the SoC with at least eight, including Sensory and Fluent, already signed up.

Long-term, “half of my algorithm partners will be from China … I think it’s just because the scale of development in China is extraordinary in hardware and software,” said Polacek.

Knowles supplies a basic toolbox of algorithms for functions such as multi-microphone beam-forming and noise suppression to get OEMs started. It expects that customers will eventually migrate to third-party software.

The SoC also includes an ARM M4 core suitable for running Linux or Android as well as management tasks. “DSPs are hard to develop on, so people may want to start with something they know … that’s the period we are in,” said Polacek.

The SoC includes 4 Mbytes of memory that can be used flexibly by any of the cores for data storage or working code. An additional 1.7 Mbytes of memory is used for dedicated purposes.

The chip can drive battery-operated systems such as rechargeable headphones, which have “the smallest [power] budgets I’ve seen,” he said. It already has design wins in systems expected to go into volume production by June.

— Rick Merritt, Silicon Valley Bureau Chief, EE Times Circle me on Google+

Related posts: